Pulsepatch

Smart wound care that catches infection early and treats it on the spot.

A wearable patch that senses infection before you can see it, scores the risk with an ML model, has an AI agent check the evidence, and treats with low-frequency ultrasound and 405 nm violet light.

Watch the 40-second film · How it works · Quick start · The math · Safety

The problem

60% of diabetic foot ulcers become infected, and 15–20% of diabetic foot infections end in amputation. Nerve damage removes pain, the body's warning sign, so infection often goes unnoticed until it's visible, and by then it's serious.

The early signs are measurable long before that. Infected wounds turn alkaline (healthy wounds sit at pH 4–6; infected ones at 7.5–9), get warmer and wetter, and bacteria use up the glucose in wound fluid. Amethyst watches for exactly those changes.

_{Sources: Diabetes Care; Nussbaum et al., Value in Health 2018. More in hardware/research/.}

How it works

flowchart LR
    P["Amethyst patch<br/>pH · moisture · 3× temp"] -->|"BLE, every 30 min"| S
    C["CGM<br/>blood glucose"] --> S
    S["FastAPI server<br/>POST /api/readings"] --> A{{"Gemini agent"}}
    A -->|tool| M["XGBoost model<br/>risk = likelihood × severity"]
    A -->|tool| D["Dosing engine<br/>tiers + safety caps"]
    A --> F["Notes, fact-checked<br/>in code"]
    S --> App["Expo app<br/>alert · therapy · trends"]
    S --> DB[("Databricks<br/>care record + dashboard")]
    D -->|"therapy command"| P

1. Sense. Every 30 minutes the patch reads wound pH, moisture (impedance) and temperature at three spots (wound edge, healthy reference skin, ambient). Blood glucose comes from the wearer's CGM.
2. Predict. An XGBoost model compares each reading with the patient's own first-day baseline and outputs a 0–100 risk score and a tier: monitor, watch, treat or intensive.
3. Reason. A Gemini agent calls the model and the dosing engine as tools, then writes a clinician-ready assessment. It covers the trend, sensor faults, signals that disagree with the model, its confidence, and whether a clinician should look. It never invents a number: risk and doses always come from our code, and its notes are fact-checked against the tool results.
4. Treat. In the treat tier the patch runs ultrasound to break up biofilm, then 405 nm light to kill bacteria. Lower tiers get gentle healing ultrasound and closer monitoring.
5. Alert. The app shows the patient their risk, trends and therapy. Every reading and assessment lands in Databricks for the care team.

Results

Validated leave-one-person-out (16 folds): every person is scored by a model that never saw their data.

We shipped XGBoost: it alarms 3.1 h earlier and is 3.8× smaller and 89× faster, which matters on a wearable. The extra false alarms are absorbed by risk smoothing and the agent's review.

Which sensors matter most (AUC of each sensor on its own): moisture 0.952 · temperature 0.933 · pH 0.917 · wound glucose 0.886 · blood glucose 0.650 · pH + temp + moisture together 0.976.

All results are on simulated data (see limits). The full method is in docs/ml-pipeline.md.

The math

Training data. No public dataset of continuous wound-sensor readings with infection labels exists, so we simulated one from 30+ published studies and real Dexcom glucose traces from 16 people (PhysioNet):

$$600 \text{ wounds} \times 14 \text{ days} \times 48 \text{ readings/day} = 403{,}200 \text{ labelled readings}$$

Features. 5 signals × (6 h median, change vs. day-1 baseline, 6 h change, 24 h change, 24 h std) + 2 time-of-day terms = 27 features.

Risk = likelihood × severity. Trained on cleanly separated data, the classifier is near-certain for any sustained shift. So we scale its probability by how far the signs have actually moved toward a full infection:

$$p = P(\text{warning}) + P(\text{infection})$$

$$s = \text{mean of the 3 largest of } \left\lbrace \frac{\Delta \text{pH}}{0.65},\ \ \frac{\Delta T}{1.6^{\circ}\text{C}},\ \ \frac{-\Delta \ln Z}{-\ln 0.65},\ \ \frac{-\Delta g / g_0}{0.40}\right\rbrace$$

Each Δ is a 2 h median minus the patient's day-1 baseline. The denominators are a full infection's average effect in the simulator: pH +0.65, temperature +1.6 °C, impedance −35%, wound glucose −40%. Each term is clipped to [0, 1.25] and s to [0, 1]. Using the top 3 of 4 means a wound whose pH doesn't rise, which happens in real infections, still scores. The score is then smoothed so one noisy reading can't trigger treatment (τ = 3 h, Δt = 0.5 h):

$$R_t = R_{t-1} + \alpha(100 p_t s_t - R_{t-1}), \qquad \alpha = \frac{0.5}{3 + 0.5} \approx 0.143$$

Tiers use 10-point hysteresis: each is entered at its threshold and left only 10 points below it, so the device doesn't flicker.

Light dose (treat tier and above) at 10 mW/cm², with the 36 J/cm² daily cap split over 3 sessions:

$$x = \text{clip}\left(\frac{R - 30}{60}, 0, 1\right), \qquad E = 12x \ \text{J/cm}^2, \qquad t = \frac{E}{0.010 \ \text{W/cm}^2}$$

Ultrasound (treat tier and above), run before the light:

$$y = \text{clip}\left(\frac{R - 55}{35}, 0, 1\right), \qquad t_{\text{US}} = \max(4,\ 10y) \ \text{min}, \qquad I = 0.1 + 0.4y \ \text{W/cm}^2$$

The AI agent

wound_agent/ wraps the model and dosing in three tools: get_sensor_history → predict_infection → plan_treatment. The LLM writes structured notes: summary, trend, reasoning, data_quality_flags, conflicting_signals, recommend_clinician_review and confidence.

• Gemini by default. Gemini 3.8 Flash with low thinking takes about 5 s and about $0.005 per assessment. GEMINI_MODEL=gemini-3.1-pro-preview with GEMINI_THINKING=high gives maximum care. OpenRouter and Claude Opus also work (AGENT_PROVIDER=gemini|openrouter|claude).
• Fact-checked. After the agent writes, code checks its notes against the tool results (risk score, tier, lifted patch, clinician review) and asks for one correction if anything is off.
• Knows what to distrust. It's told each sensor's standalone accuracy, that infections don't always raise pH, and how hysteresis works. In testing it caught a 43 °C reading as a sensor fault rather than a fever, and blamed a risk jump on a meal-time glucose spike while the wound sensors stayed healthy.
• Never blocks the demo. With no API key, the same model and dosing run and a rule-based summary stands in.

Hardware

A reusable 30 × 40 mm, 4-layer pod that clips onto a disposable hydrogel pad. Full spec →

Generated, not drawn. One Python file (design.py) generates the schematic, places every part, autoroutes all 178 connections and runs KiCad's electrical and design-rule checks in one command.

Power and heat budget (details):

• Sensing averages ~22 µA, which is about 60 days on a 40 mAh cell.
• At 10 mW/cm² of light, skin reaches 37.3 °C, or 38.7 °C on a poorly perfused diabetic foot, including the LEDs' own heat. Both stay under our 41 °C ceiling.
• Doubling to 20 mW/cm² would hit 43.4 °C, so we didn't.

Safety in three layers

Quick start

Python 3.12+ (Adam's pinned numpy 2.5 and xgboost 3.4 need it).

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env               # add GEMINI_API_KEY (or OPENROUTER_ / ANTHROPIC_API_KEY)

Run the agent on a simulated wound:

python main.py -v --scenario infected        # healthy | early_infection | infected | sensor_fault | patch_lifted
python main.py --history sample_data/simulated_wound_history.csv
python main.py --offline                     # model + dosing only, no LLM

Live demo + hardware API (then open http://localhost:8000):

python demo_ui/server.py
python demo_ui/send_reading.py --scenario infection --count 8 --every 3    # pretend to be the patch

Phone app (Expo Go):

cd app && npm install && npx expo start      # add --tunnel on campus Wi-Fi

Firmware tests (simulated board) and board build (KiCad 10):

cd firmware && make test && make arm-check
cd hardware/kicad && ./build.sh              # about 6 min with autorouting; --no-route for seconds

Databricks care record (landing from the databricks branch; design in docs/superpowers/specs/):

python main.py --databricks                  # readings + assessments to Databricks, care-team dashboard

Film:

cd remotion && npm install && npm run dev

Tests:

python -m pytest

Repo map

Honest limits

• Simulated training data. The model has never seen a real wound. Retrain on real patch data before any clinical use.
• Prototype doses from published studies, not a validated medical protocol.
• The board is designed and autorouted but not yet fabricated, and the firmware has only run on a simulated board. Light and ultrasound intensity at the wound are estimates until measured.
• Not all on the board yet. The board's ultrasound runs at 30 kHz (within the 20–40 kHz range of clinical wound devices). The 1.5 MHz healing mode and wound-fluid glucose sensing aren't on this board yet.
• Combining light and ultrasound is backed by lab (in-vitro) studies so far.

Team

Built at WolfHacks 2026 by:

• Adam Rakhmanov: ML model, simulator, dosing, demo server.
• Anay Vaidya: hardware, firmware, Databricks.
• Aarush: app and film.
• Suhaan Temkar: AI agent.

_{Glucose data: BIG IDEAs Lab Glycemic Variability and Wearable Device Data v1.1.3, PhysioNet, ODC-By 1.0.}